Sankalp Sharma
Paperwork Labs is a one-person fintech and AI platform: seven production apps, and every line of code authored, reviewed, and shipped by an AI agent fleet I designed. Before that I spent a decade on data platforms: at Cruise I built the commercial and financial data warehouses, then Medallion, one architecture unifying all company data, from road and driving data to operations, marketing, legal, and realtime services, all of it built as code with Terraform and dbt, serving 300+ data scientists and analysts and directly supporting the company's top-line metrics; at Credit Karma, recommendations reached 120M+ members.
Open to forward-deployed and AI platform engineering roles · San Francisco
Verified from the repository on 15 September 2026.
The company is a closed loop. Agents file the issues, write the code, review each other's pull requests, merge through a queue, and revert their own mistakes. I set direction and make the calls that matter; the fleet does the rest.
Provider-agnostic by contract
Two-layer routing separates what a task needs from where it runs: capability tiers pick the model class, billing lanes pick the provider and who pays. The system can move between Claude, GPT, Cursor, Codex, and Claude Code without rewriting product code, and nothing silently downgrades.
Disciplined like a team
Agents file issues, write code, review pull requests, merge through a queue, and revert mistakes. I set direction and make the calls that matter; the fleet does the rest through gates that fail closed.
Owns its infrastructure
Paperwork ID is the self-hosted authentication layer behind the product suite. Accounts owns the sign-in shell, products exchange one-time handoff tokens for scoped sessions, and backends verify JWTs through JWKS.
The decade before
Before Paperwork Labs, I spent a decade building the data platforms that operators and customers depended on every day: Cruise's commercial warehouse and the governance around it, and Credit Karma's streaming and recommendation pipelines at 120M-member scale.
Live
Portfolio intelligence and automated trading. It connects five brokerages, backtests strategies walk-forward against an independent oracle, tracks tax lots, and refuses any trade that fails its risk gates.
In IRS certification
AI tax preparation, now under certification. Photograph a W-2 and it becomes a deterministic, integer-cents return, computed by a tax engine that covers all 50 states plus D.C.
In development
Business formation end to end: filings, compliance dates, and a deterministic handler for each state portal, currently under certification.
In development
A compliance workspace for CPA firms. Document workflows stay structured, tax routes stay scoped to the firm, and every session runs through Paperwork ID.
Internal OS
The founder command center. One surface to run the products, direct the agent fleet, hold the secrets, and keep a dated record of every operating decision.
Agent runtime
The agent runtime every product shares. Personas route the work, episodic memory recalls it through RAG and hybrid retrieval on pgvector, schedules keep it moving, and MCP hands each product's capabilities to the fleet.
Feb 2026 – Present
Founded and run a multi-product fintech/AI platform: 7+ production apps, 6,800+ merged PRs, and an in-house AI agent fleet that authors, reviews, and ships the code.
2021–2026
Spearheaded Medallion, the architecture that became the source of truth for Cruise's self-driving commercial fleet: data inconsistencies down 80% for 600+ users as the team scaled from 2 to 15, scoping the platform directly with its finance, commercial, and legal stakeholders. The whole platform was code: Terraform for infrastructure, dbt for transformations.
2019–2021
Owned email recommendations for 120M+ members, contributing $20M+ in incremental monthly revenue, and built Credit Karma's first streaming pipeline.
2017–2019
Maintained LIA, a community platform serving 500+ communities and 800M members, and helped move monolith surfaces into server-rendered microservices.
2014–2017
Modernized AlphaXD, a J2EE device-monitoring platform behind 80%+ of broadband providers, cutting load times from 2 minutes to 5 seconds.
2010–2014
Helped owner-operated businesses establish a complete internet presence, translating plain-language needs into shipped websites and practical digital workflows.